Primal Recovery Strategies for Lagrangian Dual Subgradient-Based Methods